Training and Inference: Two Phases of Learning in the Age of AI
Imagine a future where everyone owns a personal tutor โ someone with the patience of a saint and the clarity of Richard Feynman โ available around the clock to guide, assess, and inspire you. That future is closer than we think.
As Andrej Karpathy popularised the idea of vibe coding, the same "vibe" concept extends naturally to learning. Why not vibe learn โ conversing with an AI tutor, exploring ideas interactively, building understanding at your own pace? It sounds wonderful. But there is a catch.
A tale of two perspectives
After the Peter Kirstein Lecture given by Prof. Tom Mitchell last year, I spoke with Dr. Yuzuko Nakamura about the adoption of AI in software engineering education. She was rather pessimistic: students are submitting coursework without solid understanding of fundamental concepts, syntax, and principles. The tool does the work; the learning doesn't happen.
Prof. Tom Mitchell, on the other hand, stayed positive โ AI empowers a personalised learning journey with rich, immediate feedback. Both are right. The question is when and how AI enters the process.
I think the answer lies in recognising that learning has two distinct phases โ and borrowing from machine learning vocabulary helps make the distinction sharp.
Phase 1: Training
The first phase is the understanding of fundamental knowledge and skills. Students need to build solid mental models before they can meaningfully leverage tools.
Think about arithmetic. You need to grasp how numbers work โ place value, operations, estimation โ before a calculator becomes a powerful instrument rather than a crutch. Think about writing. The act of thinking as you write is not equivalent to letting AI write it up for you. The struggle is the point.
In this stage, students must overcome desirable difficulty and think independently. Only by doing so will they learn how to judge assumptions, set boundaries, make rules, and nail down decisions. AI cannot learn these things for you. Going through the experiments, the analysis, the reduction, the calculations, the comparisons, and the reasoning on your own โ that is irreplaceable.
Phase 2: Inference
The second phase is applying those skills to real-world scenarios and scaling up with careful engineering. It is more about context engineering, iterative optimisation, and trial-and-error.
Take vibe coding as an example. A programmer with solid fundamentals can rapidly build a prototype and verify the ideas in their mind. In this end-to-end process, we care more about the outcome than the learning journey. Given a set of knowledge, skills, principles, boundaries, and objectives, AI agents can plan, execute, reflect, and iterate progressively. And they are getting remarkably good at it.
This is where AI belongs โ not replacing your thinking, but amplifying it. The AI in Learning newsletter captures this well: the most promising applications are the ones that meet learners after the foundations are in place.
The illusion of understanding
Here is the uncomfortable truth for learners in the age of AI: when knowledge seems more accessible, the real hard things still lie underneath.
Think twice if you find yourself understanding something effortlessly. Is it genuine comprehension โ or is it a hallucination of understanding, where AI has done the heavy lifting and you merely feel like you know the basics?
The Training phase is where the real work happens. It is slow, often frustrating, and deeply personal. No shortcut, no matter how intelligent, can substitute for it.
The Inference phase, however, is where AI becomes your superpower โ if you have earned the right to wield it.
Inspired by Prof. Tom Mitchell's Peter Kirstein Lecture at UCL and the AI in Learning February 2026 newsletter.